Research map / Architectures and efficiency / Attention and transformers
Block-State Transformers
Jonathan Pilault, Mahan Fathi, Orhan Firat, Christopher Pal, Pierre-Luc Bacon, Ross Goroshin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| place single BST layer at index 9 instead of 3 (ablation) | BST:SH:S4 · PG19 | perplexity: 12.41 -> 11.88 | Improved |
| single BST layer at index 12 instead of 9 (ablation) | BST:SH:S4 · PG19 | perplexity: 11.88 -> 12.03 | Hurt |
| increase number of BST layers from 2 to 4 (ablation) | BST:SH:S4 · PG19 | perplexity: 11.69 -> 11.21 | Improved |
| increase number of BST layers from 4 to 5 (ablation) | BST:SH:S4 · PG19 | perplexity: 11.21 -> 11.20 | No clear effect |
| increase S4 state size from 8 to 16 (ablation) | BST:SH:S4 · PG19 | perplexity: 11.95 -> 11.57 | Improved |
| increase S4 state size from 16 to 64 (ablation) | BST:SH:S4 · PG19 | perplexity: 11.57 -> 11.54 | No clear effect |
| use BST:MF instead of BST:SH (S4) | BST:SH:S4 · arXiv | perplexity: 2.51 -> 2.48 | Improved |
| use BST:SH instead of BST:MF (S4) | BST:SH:S4 · PG19 | perplexity: 11.63 -> 11.57 | Improved |
| use unstructured filters instead of S4 (SH) | BST:SH:S4 · PG19 | perplexity: 11.57 -> 11.52 | Improved |
| BST:SH:S4 versus Mega-chunk | BST:SH:S4 · LRA Avg | accuracy: 85.66 -> 86.96 | Improved |
| BST:SH:S4 versus BRecT | BST:SH:S4 · LRA Avg | accuracy: 60.80 -> 86.96 | Improved |
| BST:SH:S4-L versus GSS-Hybrid-L | BST:SH:S4 · PG19 | perplexity: 10.52 -> 10.47 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Mega — Mega: Moving Average Equipped Gated Attention (ICLR 2023) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality (ICML 2024 · similarity 0.89)
- Block Transformer: Global-to-Local Language Modeling for Fast Inference (NeurIPS 2024 · similarity 0.88)
- The Expressive Capacity of State Space Models: A Formal Language Perspective (NeurIPS 2024 · similarity 0.88)
- Block-Biased Mamba for Long-Range Sequence Processing (NeurIPS 2025 · similarity 0.87)
- Compress, Gather, and Recompute: REFORMing Long-Context Processing in Transformers (NeurIPS 2025 · similarity 0.87)
- Transformer Block Coupling and its Correlation with Generalization in LLMs (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
